[3B4-GS-11] AI and Society: medicine

[3B4-GS-11-01] A feasibility study on lightGBM for predicting psychological distress scale of remitted MDD patients from life-log data

〇Ayano Hata¹, Ryo Kiguchi¹, Yuuki Yoshida¹, Yoshitake Kitanishi¹, Junichiro Yoshimoto², Aran Tajika³, Toshiaki Furukawa³ (1. SHIONOGI & CO.,LTD., 2. Nara Institute of Science and Technology, 3. Kyoto University)

[[Online]]

Keywords:lightGBM, Depression, recurrence prediction

Depression is a psychiatric disorder with a high prevalence and a high recurrence rate. It is also known that the higher the number of recurrences, the higher the probability of future recurrences. In order to prevent recurrence, early detection of warning signs of recurrence is very important. Recent advance in wearable devices allows us to continuously and non-invasively collect user’s activity log and biometric data such as sleep, and heart rate. We developed an algorithm for predicting recurrence of depression by creating features from the activity log and biometric data, and applying ensemble learning with lightGBM. The results showed the algorithm could predict K6 score (an index of psychological distress ranging from 0 to 24 points) with a practical performance (coefficient of determination >= 0.7 and mean absolute error <= 1.7). Here, we discussed and reported the factors that contributed to the improvement in accuracy. The study was conducted as the second analysis of the Fun to Learn, Act and Think through Technology-2 (FLATT2) study.